Remaining useful life prediction of lithium-ion battery using a hybrid model-based filtering and data-driven approach

Abstract

Lithium-ion battery has been widely used for industrial systems, however, its performance gradually deterioration with cycling is an inevitable problem. The main reason for unreliable remaining useful life (RUL) prediction is that the adaptability and accuracy of the physical degradation models are not fully considered. In this paper, a new hybrid algorithm using a combination of model-based Kalman filtering (KF) and data-driven relevance vector machine (RVM) is proposed to track batteries' degradation trend. The data-driven RVM is employed to learn the capacity degradation trend and then predict multi-step future observation series, which are fed into KF as new observations. Also, since the RVM implements prediction through iterative calculation, its accumulative error is always being neglected. A weighting function is designed with the purpose of adjusting the updating step in KF in order to avoid over-dependency on either future observations or degradation model, which realizing a reasonable balance between system model and future observations. Simultaneously, considering the system dynamics may change during the prediction period, the modified Sage-Husa adaptive filtering algorithm is utilized to update the observation noise covariance dynamically. Finally, the feasibility and validation of the proposed method are examined using the batteries test data from NASA PCoE.